# Full text of "Smooth neighbors"

## See other formats

SMOOTH NEIGHBORS J. B. CONREY, M. A. HOLMSTROM, T. L. MCLAUGHLIN Abstract. We give a new algorithm that quickly finds smooth neighbors. I. Introduction We say that a number is z-smooth if none of its prime factors exceed z. In this paper we search for solutions 6 of p | 6(6 + 1) p < z. (1) If 6 is a solution of (OQ), then we refer to the pair (6,6 + 1) as ^-smooth neighbors. It has been known since the work of Stormer in 1898 [S] that for any z there are only finitely many ^-smooth neighbors. In 1964, D. H. Lehmer [LI] found all 869 of the 41-smooth neighbors. To do this he improved upon Stormer's method, which relied on solving a finite number of Pell's equations. In fact Lehmer showed that if b(b + 1) is ^-smooth then 46(6 + 1) is the u y part" of the nth power of the fundamental solution xq + y/dyo of the Pell's equation x 2 - dy 2 = 1 where d is squarefree and z-smooth, and where n < (d + l)/2. Lehmer solved all of these Pell's equations to see which led to 41-smooth neighbors. In 2011 Luca and Najman [LN] used a modified version Lehmer's method to find 100- smooth neighbors. In a calculation that took 15 days on a quad-core 2.66 GHz processor, they found 13,325 neighbors and claimed that this was all of the possible 100-smooth neighbors. The calculation was especially difficult because solutions to Pell's equations for squarefree 100-smooth integers can have as many as 10 10 digits. Consequently they had to use a special method to even represent the solutions. In an erratum [LN1], they found 49 more solutions that they had missed previously. We have found a fast, amazingly simple algorithm that finds almost all z-smooth neighbors much more quickly. In fact, when we ran our method to find 100-smooth neighbors, it completed in 20 minutes (on a similar machine as Luca and Najman used) and found 13,333 100-smooth neighbors. We were missing 37 solutions that Luca and Najman found. Subsequently we searched for all 200-smooth neighbors. This computation took about 2 weeks and produced a list of 346,192 solutions. This list included all but one of the solutions from the (completed) Luca-Najman list. We determined that this 100-smooth number missing from our list would be found using our method when searching for 227- smooth neighbors. l 2 J. B. CONREY, M. A. HOLMSTROM, T. L. MCLAUGHLIN A plot of the logarithms of the first members from our list of 200-smooth neighbors suggests that they are normally distributed. 2. The algorithm Suppose we have a set S of positive integers. For any two elements b < B of S form the ratio A h B + 1 J' ~ 6+1 X B where gcd(/3, (3') = 1. Sometimes it will be the case that 0' — (3 + 1, for example 15 _ 3 5 16 ~ 4 X 4' We are particularly interested when this happens, i.e. we are interested in the solutions of P + 1 b+ 1 B K ' where b and B are in 5*. Given S, we form a new set, S', which is the union of 5* and all of the solutions (3 to §Z§. We can repeat this process to form (S')' = S" and so on. Ultimately, by Stormer's theorem[S], we will arrive at a set (meaning n iterations of priming) for which (S^Y = S( n \ i.e. there are no new solutions to be added. We let 5(S) denote this set that can no longer be enlarged by this process. As an example, suppose that S = {1,2,3,4,5}. Then, it is easy to check that S' = {1,2,3,4,5,8,9,15,24}. and S" = {1,2,3,4,5,8,9,15,24,80}. After that we have 5"" = S", so £({1,2,3,4,5}) = {1,2,3,4,5,8,9,15,24,80}. Recall that Lehmer [LI] gave a complete list of 41-smooth solutions to ([T|). In particular, we can see from his list that the above set {1,2,3,4,5,8,9,15,24,80} is the complete list of 5-smooth solutions to (CE]). In other words, we found all of the z = 5 solutions to §B) by starting with the set {1, 2, 3, 4, 5} and then repeatedly adding in solutions to (T5]). This good fortune is not always the case. For example, 5({1, 2, 3, 4, 5, 6, 7}) = {1,2, 3, 4, 5, 6, 7, 8, 9, 14, 15, 20, 24, 27, 35, 48, 49, 63, 80, 125, 224, 2400} whereas from Lehmer's table we see that the complete set of 7-smooth solutions to (OQ) includes all of these numbers along with 4374. However, it is the case that 4374 e {1,2,3,4,5,6,7,8,9,10}. SMOOTH NEIGHBORS 3 Actually, 8({1, 2, 5, 6, 10}) = {1,2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 14, 15, 20, 21, 24, 27, 32, 35, 44, 48, 49, 54, 55, 63, 80, 98, 99, 120, 125, 175, 224, 242, 384, 440, 539, 2400, 3024, 4374, 9800} contains all of the solutions to ([[]) with y = 7. It is easily checked that 2 /r,\ -2 /r \ -4 4374 4375 so that the fraction yy, which corresponds to b = 10, does not appear; but in finding the above decomposition by tracking the iterative procedure in computing 6({1, 2, 5, 6, 10}) this fraction made an appearance and then is later canceled out. Looking at Lehmer's tables we find for all primes p up to 41, with the exception of p = 7 and p — 41 that z p :=S({l,2,...,p}) coincides exactly with the complete set of p-smooth solutions of ([1]). In the case p = 41, we found that Z41 has 890 of Lehmer's solutions; it is missing the largest solution, 63927525375. Note that 63927525375 3 3 x 5 3 x 7 7 x 23 The least n for which 63927525376 2 13 x ll 4 x 13 x 41' 63927525375 E 5{{1, 2, . . . , n}) is n = 52. We calculated Zigg in a week on a PC using Mathematica. It has 346192 elements. A histogram of the logarithms of these numbers is shown in Figure 1. The prime before 199 is 197. The number of new b in z 19 g, i.e. those not in z 197 , is 43215. For all but 300 of these b it is the case that 197 | b{b + 1). The other 300 have a smaller largest prime factor. Here is a table of the number of these b sorted by the largest prime factor of b{b + 1): Prime 127 131 137 139 149 151 157 163 167 173 179 181 191 193 Number 1 1 2 2 5 5 10 8 13 16 25 36 43 51 197 82 3. Solving the Diophantine equation In this section we analyze the solutions of b B + l /3 6+1 X B ~ /3 + 1 This equation can be written as 6(fl + l)(/9 + l) = {b+l)B(3. 4 J. B. CONREY, M. A. HOLMSTROM, T. L. MCLAUGHLIN 4000 3000 2000 1000 10 20 30 40 Figure 1. A histogram of the set {log b : p | b(b + 1) with the PDF of the normal distribution. 50 p < 199} together 40 000 30000 20000 10 000 10 20 30 40 Figure 2. A plot of (n, \z Pn \ — |-2 P „_J) where p n denotes the nth prime. Let gcd(6, B) = g and put b = gu B = gv where clearly gcd(w,f) = 1. Then we have u(B + l)(/3 + l) = (b + l)v(3. Now, gcd(w, (6 + l)v) = 1 so it must be the case that u \ {3 , say (3 = hu. Then we have (5 + 1)03 + 1) = (b + l)vh. SMOOTH NEIGHBORS 5 Then v \ (f3 + 1), say (3 + 1 = vx and h \ (B + 1), say B + 1 = /«/. Thus, in terms of the variables g, h, u, v, x, y, we have the three equations hy - gv =1 vx — hu =1 xy — gu =1 along with b = gu, B = gv and (3 = hu. If we solve for x and y in the first two equations and substitute the results into the third equation we have uh + gv = hv — 1. (3) Notice that if h and v are given positive coprime integers then there exist unique positive integers u and g that satisfy ([3]). Then x and y can be determined as well as b,B and (3. Thus, a coprime pair (h, v) leads to a triple (6, B, (3) and vice- versa. From another perspective, one may ask how to find a pair (6, B) from a given (3. Take any divisor w of (3 and any divisor v of /3 + 1. Then /i = and x = (/3 + l)/f, from which we see that v — u h — x y = T~ 9 vx — hu vx — hu Thus, the pair (u, v) leads to the pair (b, B) given by uh — ux n u . n , . „ hv — xv v n . n „ . f 6 6= — =(3 --((3 + 1) B = — = -13- {(3 + 1 ) = -, vx — hu v vx — hu u u where we assume that u < v. And we've already seen that given b < B we can define b B u (b,B) (b, B) ' Thus, given (3, there is a one-to-one correspondence between pairs (u, v) with u < v, u \ (3, v | (/3 + 1) and pairs (6, 5) with b < B and x = -^j. 4. Trees In this section we illustrate the way in which 601425 enters into Z20 = <5({1, 2, . . . 20}). 6 J. B. CONREY, M. A. HOLMSTROM, T. L. MCLAUGHLIN This tree illustrates the factorization 601425 601426 6\ [7\ 1 /8\ Via /i5\ fn\ 1 /is\ /19 7J\8J \9 J \12J\1QJ\18J V 19 7 V 20 5. LUCA AND NAJMAN'S WORK Luca and Najman published a paper in Mathematics of Computation in 2011 which im- proved on Lehmer's previous work [LN]. They gave a list of 13325 pairs of numbers which were 97-smooth neighbors and claimed that this list was complete. Their computation took 15 days on a computer with a 2.5 GHz processor. By comparison, when we searched for 97-smooth neighbors, our method took just under 20 minutes and, much to our surprise, produced 13333 neighbors. In comparing their list with ours, Luca and Najman found 37 solutions not on our list. However, we found 45 solutions not on their list! Independently of our work, Luca and Najman found the 49 solutions they had missed previously. Our calculations of 199-smooth numbers took more than a week and produced 345192 solutions, some of which were the solutions we had missed in our 97-smooth calculation. In fact, our ^199 is only missing one solution from the Luca-Najman list, 9591468737351909375. 6. The largest Luca-Najman solution The 79-smooth neighbor pair which starts with (3 = 9591468737351909375 was found by Luca and Najman but is not on our list (which we call z 199 ) of 199-smooth neighbors. We can prove that it will first appear in our method when we search for 227- smooth neighbors. Given a number (3, there is a 1-1 correspondence between pairs (b, B) with b(B + l)/((6 + l)B) = (3/((3 + 1) and pairs (u,v) with u \ (3 and v \ ((3 + 1). In one direction this correspondence is given by a b = 0--O3 + 1) B = ^-0 + 1) = !* U U There are 1440 divisors of (3 and 5632 divisors of (3 + 1. This means that there are a total of 8110080 pairs (u,v) to consider. For each of these pairs we computed the pair (b,B) and then computed the largest prime factor of b(b + 1)B(B + 1). The least out of all of these SMOOTH NEIGHBORS 7 largest prime factors was p = 227 which appeared for several pairs (b,B), in particular for the pair b = 285406166331883519 B = 294159243066390624. Therefore, /3 cannot appear as a solution in our method unless we go up to p = 227. Further analysis led us to find an explicit tree for /3 all of whose bottom nodes are either in ^199 or else appear within the first two iterations (which are quick to compute) arising in the computation of z 2 27', thus we can show that f3 G ^227 without computing all of ^227. We now describe the tree in detail. Given a (3 we have described a 1-1 correspondence between pairs (u,v) and pairs (b,B). For a pair (b,B) let us use the notation F(b,B)=P to denote that b (£ + !) _ g (6+1) X B ~(/3 + l)' Thus, F(285406166331883519, 294159243066390624) = 9591468737351909375. is the first step of our computation. Let's use b = 285406166331883519; B = 294159243066390624; (3 = 9591468737351909375 for the rest of this section, so that F(b, B) = (5. Now let gi = 2229716045541599; g 2 = 2247272709023744 h x = 186642247267999999; h 2 = 510640590102749183. F( gi ,g 2 ) = b; F(h 1 ,h 2 ) = B. and Then Moreover, 9l £ ^199 h 2 G ^199 so that we may restrict our attention now to g 2 and hi. Next, we let i t = 907177810312319; i 2 = 911608699868750 and ji = 1671690051584; j 2 = 1672934505788. Then F(i 1 ,i 2 ) = h 1 F(j 1 ,j 2 )=g 2 and j 2 G ,2199; we are left to decompose ii,i 2 , and j\. Let h = 341611712; k 2 = 341681535; m 1 = 300775; m 2 = 301040. These satisfy F(ki, k 2 ) = ji F(mi, m 2 ) = k 2 8 J. B. CONREY, M. A. HOLMSTROM, T. L. MCLAUGHLIN and ki,m 2 G £199. Let ni = 3405; n 2 = 3444; o x = 454; o 2 = 524; then F(n 1 ,n 2 ) = mi, F(o 1 ,o 2 )=n 1 and n 2 ,o 2 G ,2199. Let Pi = 199802399641; p 2 = 199846415040; q x = 16503580; q 2 = 16504943; then F(p 1 ,p 2 ) = i ± ] F(q 1 ,q 2 ) = p 2 ; and p x G 2:199. Let n = 561824; r 2 = 581624; Sl = 49664; s 2 = 54480; then F(r 1 ,r 2 ) = q ± ; F(si,s 2 ) = r±; and r 2 , Si G 2199. Let h = 6810; t 2 = 7783; u x = 108033083250000; u 2 = 122557101693480; then F(t 1 ,t 2 ) = s 2 ; F(u l7 u 2 ) = i 2 ; and t 2 ,Ui G z 199 . Let v l = 10638314820; v 2 = 10639238337; w 1 = 1451240; w 2 = 1451438 then F(vi,v 2 ) = u 2 \ F(wi,w 2 ) = v\\ and v 2l w\ G Z199. Let x 1 = 92852; x 2 = 99198; y x = 4312; y 2 = 4508; then F(x 1 ,x 2 ) = w 2 ; F(y l ,y 2 ) = x 2 ; and x X) y 2 G z 199 . Finally, o 1 ,t 1 ,y 1 G z 227 . This last fact is determined by computing the first two iterations in the process for determining 2227 which only takes a few seconds. 7. Computational time We have computed 346192 solutions to p I 6(6 + 1) p < 199. 199 is the 46th prime. Would it be possible to use the Pell's equation method to find the complete list of 199-smooth neighbors? No. To use the Pell's equation method [LI] to find all of the solutions, one would first have to find the fundamental solutions of 2 46 - 1 = 70368744177663 different Pell's equations and then check up to the 100th solution of each. The difficulty in finding the continued fraction expansion of \fd grows dramatically with d. We expect that for the prime 199 we would find d such that the period of \fd is as large as 10 40 . Even using the subexponential algorithm employed by Luca and Najman [LN] it would still be impossible to go this far. Would it be possible by this method to find all of the solutions we found? No. For the numbers b on our list £199 the continued fraction expansions of \/b(b + 1) have unusually small periods; in fact the largest such period is only 38. Suppose one does a calculation where they check all 2 46 square-free gPs that are 200-smooth to see which lead to short continued SMOOTH NEIGHBORS 9 fraction expansions as described above; even this calculation we estimate would take, at a conservative minimum, at least 3000 years on the computer we used. 8. Largest solutions Here is a list of (q, 6) where q is a prime number up to 197 and 6 is the largest element of Z199 for which q | (6(6 + 1) and p | 6(6 + 1) =>- p < q. (2, 1), (3, 8), (5, 80), (7, 4374), (11, 9800), (13, 123200), (17, 336140), (19, 11859210), (23, 5142500), (29, 177182720), (31, 1611308699), (37, 3463199999), (41, 63927525375), (43, 421138799639), (47, 1109496723125), (53, 1453579866024), (59, 20628591204480), (61, 31887350832896), (67, 12820120234375), (71, 119089041053696), (73, 2286831727304144), (79, 1383713998733898), (83, 17451620110781856), (89, 166055401586083680), (97, 49956990469100000), (101,4108258965739505499), (103, 19316158377073923834000), (107, 386539843111191224), (109, 90550606380841216610), (113, 205142063213188103639), (127, 20978372743774437375), (131, 1043073004436787852800), (137, 65244360004072055000), (139, 152295745769656587384), (149, 6025407960052311234299), (151, 1801131756071318295624), (157, 277765695034772262487), (163, 1149394259345749379424), (167, 2201197005772848768608), (173, 4574658033790609920000), (179, 9021820053747825025975), (181, 13989960217958128903124), (191, 75121996591287627735039), (193, 444171063468653314858175), (197, 25450316056074220028640), (199, 589864439608716991201560) Note that the Luca-Najman number 9591468737351909375 would correspond to q = 79 above, where we have 1383713998733898. 9. Smoothness and the ABC-conjecture The ABC equation is A + B = C. (4) The 1985 ABC-conjecture of Masser and Oesterle asserts that for any e > there is a k(c) > such that for all solutions to (j4]) the inequality C < /?(e)rad(ABC) 1+e holds, where rad(n) := n p |nP> caue d the radical of n, is the product of the prime divisors of n. In studies of the ABC conjecture solutions to (HI) there are a variety of indicators used 10 J. B. CONREY, M. A. HOLMSTROM, T. L. MCLAUGHLIN to measure the solutions. A chief one is the quality: logC q(A,B,C) lograd(ABC)' In fact, it seems to be standard to call any triple (A, B, C) whose quality is at least 1 "an ABC-triple." Much work has been done to systematically find all ABC-triples. The first 22763667 ABC-triples are available for download from the abc@home web-site [A]. These include all ABC-triples with C < 10 18 . Lagarias and Soundararajan [LS] define what they call the smoothness exponent k of an ABC solution by k {A,B,C) = - — - — - log log C where z is the largest prime factor of ABC, and they define the number kq by k,q := liminf Kq(A, B, C). C— >oo They observe that the ABC conjecture implies that k,q > 1 and conjecture that k = 3/2. They also prove that the Generalized Riemann Hypothesis implies that Kq < 8. We introduce another measure of the size of a z-smooth solution of with a quantity we call the smoothness index given by S( A,B,C).J^ log z where we take z to be the largest prime factor of ABC . For example, 2 + 25 = 27 has C = 27 and the largest prime factor of 2 x 25 x 27 = 2 x 3 3 x 5 2 is 5 so that s(2, 25, 27) = M = 2.04782. log 5 Remark 1. For comparison, it is easy to make the smoothness exponent large (just take C to be a large prime ) but hard to make it small. For the smoothness index, the challenge is to make it large. Remark 2. The quality is more a measure of the average size of the prime divisors of ABC whereas the smoothness index measures the size of the maximum prime divisors of ABC . Note that an ABC-triple can have a high quality even if a few of the prime divisors are large whereas that cannot happen for measures of smoothness. Remark 3. The ABC conjecture asserts that q — > 1 as C — > oo. However, by work of Balog and Wooley [BW] s(A,B,C) is unbounded. Remark 4. Work of Lagarias and Soundararajan [LS], leads us to expect triples with s(A,B,C) about as large Thus, theoretically we know s(A,B,C) can be arbitrarily large; for us the challenge is actually finding (A, B, C) with s(A, B, C) large. In other words it's a computational challenge SMOOTH NEIGHBORS 11 rather than a theoretical challenge to find large values of s(A,B,C). Our algorithm finds large values of s(l,B,B + l). The largest value we found is for B = 19316158377073923834000 we have s(l,B,B + l) = 11.0719 The value of B here is ~ 1.9 x 10 22 which is beyond the range where systematic study of the ABC conjecture has been conducted. We know of no larger values of s. 10. Some data about smoothness We call a triple (A,B,C) maximally smooth if s{A 1 ,Bi,C\) < s(A,B,C) for all C± < C (or for Ci — C and A\ < A). Here are the first thirteen maximally smooth triples, together with their smoothness index: 1 + 3 = 4 1.262 3 + 5 = 8 1.292 1 + 8 = 9 2 2 + 25 = 27 2.048 5 + 27 = 32 2.153 1 + 80 = 81 2.730 3 + 125 = 128 3.015 32 + 343 = 375 3.046 49 + 576 = 625 3.308 5 + 1024 = 1029 3.565 1 + 2400 = 2401 4 1 + 4374 = 4375 4.308 7168 + 78125 = 85293 4.427 The triples on this list seem very appealing. In particular, 3 + 5 3 = 2 7 1 + 5 x2 4 = 3 4 2 5 + 7 3 = 3 x 5 3 5 + 2 10 = 3 x 7 3 1 + 3 x 2 7 = 7 4 7 x 2 10 + 5 7 = 13 x 3 8 12 J. B. CONREY, M. A. HOLMSTROM, T. L. MCLAUGHLIN are all of the form ax 1 + by m = cz n with a, b, c, x, y, z each being 1 or a prime and £,m,n > 2. A triple (A, B, C) is called an "ABC-triple" provided that C > rad(ABC). The first 22763667 triples are available for download from the abc@home web-site [A]. These include all of the triples for C < 10 18 . The largest C for a triple on this list is for the triple 131854322526743011 + 9091517323167918864 = 9223371645694661875 or in factored form 13 4 x 16651 3 + 2 4 x 3 8 x 53 4 x 3313 2 = 5 4 x 7 7 x 37 x 59 2 x 373 2 . This triple has smoothness index 4.493. Here we have C = 9.2 x 10 18 . Among these more than 22 million ABC - triples the triple with largest smoothness index is 176202799695875 + 3178472661789594624 = 3178648864589290499 with smoothness index 11.0653. In factored form this is (5 3 x 23 3 x 42 5 ) + (2 10 x 3 7 x 7 6 x 13 3 x 17 2 x 19) = ll 4 x 31 x 37 4 x 43 3 x 47. 11. Lehmer's table For each prime z < 41 and each k < 6 Lehmer [L2] found the largest n such that p | n(n + 1) . . . (n + h — 1) =>- p < z. For example, for z = 37 and h = 3 he found n = 17575 which means that 17575 x 17576 x 17577 = 2 3 x 3 4 x 5 2 x 7 x 13 3 x 19 x 31 x 37 has no prime factor larger than 37 and that 17575 is the largest number with that property. Our algorithm produces many more such data points, but in each case we can only say that our number is a lower bound for the actual number, since we don't know for any z > 41 whether our list of b is complete. But to give some examples, we found that £(134848 x 134849 x 134850) = 43 £(192510 x 192512 x 192512) = 47 £(1205644 x 1205645 x 1205646) = 53 SMOOTH NEIGHBORS 13 Below is a table of lower bounds. p h — 9 11 — O h — A lb — ^ h — ^ h — fi lb — U h — 7 lb — ( q o Q o c oU Q O 7 / /I Q7/1 4o (4 4o 11 osnn yoUU c/i 04 1 Q lo IZoZUU oOU OO O/I Z4 1 7 1 / OO014U /i /in 44U OO /I s 4o 1 O iy 1 1 QCQQ1 fl iiooyziu Z4oU lOo A fi 4o OQ Zo iiooyziu Z4oU QOO oZZ /I fi 4o zy 1 771 80700 Hi loZ i ZU loom OZZ KA 04 ol loiioUooyy 1 Q/i c;/i 10404 lOlo i c;o 10Z Q7 O / o4ooiyyyyy i 7C7C 1/0/0 lOlo 1 ^O 10Z /1 1 41 ooyz / ozoo / o oi oqcn ZlZooU lOoU i ci 7 101 / Ofi K. Zoo /I Q 4o /lOI 1 QS7QQKQQ 4ziioo / yyooy oi oqcn ZlZooU lUo /o i ci 7 101 / Q/i n o4U A 7 4 / 1 1 no/iQK79Qi oc; iiuy4yo / zoizo ZlZooU 1/0/0 i ci 7 101 / Q/i n o4U 1 QA lo4 53 1453579866024 1205644 17575 1517 340 184 59 20628591204480 1205644 17575 1517 528 527 61 31887350832896 1205644 17575 1767 528 527 67 31887350832896 2018978 17575 5828 1271 527 71 119089041053696 3939648 17575 5828 1271 527 73 2286831727304144 3939648 70224 5828 1271 527 79 2286831727304144 15473808 70224 5828 3476 527 83 17451620110781856 15473808 97524 5828 4897 4896 89 166055401586083680 407498958 97524 5828 4897 4896 97 166055401586083680 407498958 97524 7565 7564 4896 199 589864439608716991201560 768026327418 61011223 1448540 44250 18904 We note some minor errors in Lehmer's tables: his h = 5 entry for p = 41 should be 1517. Also, his Table IIA is missing the entries 10935, 12901781, and 26578125 in the 29 column and 4807, 12495, 16337, 89375 from the 31 column. 12. Future directions The unanswered question in all of this is why does it work? Can we really get all of the b for which b(b + 1) has only small prime factors by repeatedly applying 5 to a small initial set? A reasonable thing to do would be to run this for a year on a supercomputer and generate millions of smooth neighbors, perhaps some with as many as 30 digits. It is true that computations get longer as you raise your smoothness barrier. But the length of computation time is due mainly to the vast number of solutions, not the complexity in finding them. 14 J. B. CONREY, M. A. HOLMSTROM, T. L. MCLAUGHLIN The same technique we have elucidated above works surprisingly well to find solutions to the more general problem p \ b(b + k) =>- p < z when k is an odd integer. Fix a difference k. Start with an initial set S and let then S" = (S')' and so on. Iterate this procedure until it stabilizes and call the result 5 k (S). It would be interesting to do extensive computations with a range of differences k. For even integers the process has to be modified somewhat. For example, to generate solutions of p | b{b + 2) =>- p < z one should start with the set z p = <5i({l, 2, . . . ,p}) for some p and then look for solutions of b B + l g 6+1 X B ~ /3 + 2 with b,B G z p . References [A] ABC@home, abcathome.com. [BW] Balog, Antal; Wooley, Trevor D. On strings of consecutive integers with no large prime factors. J. Austral. Math. Soc. Ser. A 64 (1998), no. 2, 266-276. [LS] Lagarias, Jeffrey C; Soundararajan, Kannan Smooth solutions to the abc equation: the xyz conjecture. J. Thor. Nombres Bordeaux 23 (2011), no. 1, 209-234. [LI] D. H. Lehmer, On a problem of Stormer. Illinois J. Math. 8 1964 57-79. [L2] D. H. Lehmer, The prime factors of consecutive integers, The American Mathematical Monthly, 72 (1965), 19-20. [LN] Luca, Florian; Najman, Filip On the largest prime factor of x 2 — 1. Math. Comp. 80 (2011), no. 273, 429-435. [LN1] Luca, Florian; Najman, Filip Errata to "On the largest prime factor of x 2 — 1;" to appear in Math. Comp. [S] C. St0rmer, Sur unc equatione indeterminee, C. R. Pris, 127 (1898), 752-754. American Institute of Mathematics, 360 Portage Ave, Palo Alto, CA 94306